g++10 versus g++9

Discussion of chess software programming and technical issues.

Moderators: hgm, Dann Corbit, Harvey Williamson

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
Post Reply
User avatar
xr_a_y
Posts: 1382
Joined: Sat Nov 25, 2017 1:28 pm
Location: France

g++10 versus g++9

Post by xr_a_y » Wed Jan 13, 2021 1:04 pm

is -45Elo for Minic ...

Does this happened to someone else ?
Any countermeasure (apart from falling back to gcc9)?

Thanks for your inputs !

Pi4Chess
Posts: 220
Joined: Mon Nov 16, 2020 11:13 am
Full name: Manuel Rivera

Re: g++10 versus g++9

Post by Pi4Chess » Wed Jan 13, 2021 1:17 pm

Ouch :shock: :?

Ras
Posts: 1686
Joined: Tue Aug 30, 2016 6:19 pm
Full name: Rasmus Althoff
Contact:

Re: g++10 versus g++9

Post by Ras » Wed Jan 13, 2021 2:18 pm

xr_a_y wrote:
Wed Jan 13, 2021 1:04 pm
is -45Elo for Minic ...
How many games?
Does this happened to someone else ?
No, just checked at a thousand games. However, the GCC-10 generated binary is 4.5% faster for me.
Any countermeasure (apart from falling back to gcc9)?
Run your engine with a single worker thread. Use a few positions to calculate to some depth so that it takes at least 10 seconds. Check whether reported PV and node counts match for the binaries generated with GCC-9 and GCC-10. It's not uncommon for bugs to only manifest upon a compiler update.
Rasmus Althoff
https://www.ct800.net

User avatar
xr_a_y
Posts: 1382
Joined: Sat Nov 25, 2017 1:28 pm
Location: France

Re: g++10 versus g++9

Post by xr_a_y » Wed Jan 13, 2021 6:22 pm

Ras wrote:
Wed Jan 13, 2021 2:18 pm
xr_a_y wrote:
Wed Jan 13, 2021 1:04 pm
is -45Elo for Minic ...
How many games?
Does this happened to someone else ?
No, just checked at a thousand games. However, the GCC-10 generated binary is 4.5% faster for me.
Any countermeasure (apart from falling back to gcc9)?
Run your engine with a single worker thread. Use a few positions to calculate to some depth so that it takes at least 10 seconds. Check whether reported PV and node counts match for the binaries generated with GCC-9 and GCC-10. It's not uncommon for bugs to only manifest upon a compiler update.
1000 games, so accurate enough to be a real issue.

No difference in terms of nodes count (or any other statistic Minic is able to print), so no obvious bug.
But 10% lower nps.

Anyway Minic was always slow and affected by small changes. I guess there are too much bad design that the compiler cannot always handle easily and optimize. It's propably more on my side than one gcc.

mar
Posts: 2260
Joined: Fri Nov 26, 2010 1:00 pm
Location: Czech Republic
Full name: Martin Sedlak

Re: g++10 versus g++9

Post by mar » Wed Jan 13, 2021 6:52 pm

xr_a_y wrote:
Wed Jan 13, 2021 6:22 pm
No difference in terms of nodes count (or any other statistic Minic is able to print), so no obvious bug.
But 10% lower nps.
10% lower nps - that's at most 10 elo at hyperbullet, realistically about 5.
so I wonder where you get -45 :shock: I suspect an issue with testing.
Martin Sedlak

syzygy
Posts: 4890
Joined: Tue Feb 28, 2012 10:56 pm

Re: g++10 versus g++9

Post by syzygy » Wed Jan 13, 2021 9:54 pm

xr_a_y wrote:
Wed Jan 13, 2021 1:04 pm
is -45Elo for Minic ...

Does this happened to someone else ?
Any countermeasure (apart from falling back to gcc9)?
Add - O3 to the compiler flags. (Hopefully this is a stupid suggestion.)

Post Reply